Skip to content

Planning Taxonomy and Trajectory Generation

Planning Taxonomy and Trajectory Generation curated visual

Visual: layered planning stack from route to behavior to motion to trajectory validation to controller/runtime assurance.

Autonomous planning is a layered decision pipeline. Different stacks use different names, but the same responsibilities keep appearing: choose a route, choose a maneuver, generate feasible motion, assign speed over time, validate safety, and hand an executable trajectory to control.

This page provides shared vocabulary for road AVs, indoor robots, outdoor industrial vehicles, and airport airside autonomy.


1. AV, Indoor, Outdoor, and Airside Relevance

DomainPlanning emphasisExample
Road AVRoute graph, traffic rules, lane changes, merges, intersections, comfort, prediction-aware motion.Behavior planner selects yield/creep/turn while motion planner produces a collision-free trajectory.
Indoor AMR / forkliftWarehouse graph, aisle constraints, pallet/dock interaction, pedestrian safety, fleet traffic management.Planner reserves a narrow aisle and generates a reverse docking trajectory.
Outdoor yard / mine / campusLarge site routes, trailer/container/task constraints, uneven surfaces, mixed manual/autonomous traffic.Yard tractor chooses a trailer move route and slows for blind corners and poor GNSS.
Airside AVStands, service roads, taxiway crossings, aircraft priority, jet-blast/FOD zones, ground-control instructions.Vehicle routes from baggage hall to stand, yields to aircraft movement, and enters a docking controller near the aircraft.

2. Layered Planning Vocabulary

LayerQuestion answeredTypical inputTypical outputRate
Mission / task planningWhat job should this vehicle do?fleet requests, schedule, operator commandstask assignment and goalseconds to minutes
Route planningWhich corridor or graph path reaches the goal?HD map, routing graph, closures, traffic rulesroute or lane/area sequence0.1-1 Hz
Behavior planningWhat maneuver should happen now?route, scene, prediction, rules, right-of-waymaneuver state and constraints1-10 Hz
Motion planningWhat geometric path or trajectory avoids obstacles?behavior goal, drivable area, obstaclescandidate path or trajectory5-20 Hz
Speed planningHow fast along the path?curvature, obstacles, comfort, signals, stop linesvelocity/acceleration profile5-20 Hz
Trajectory validationIs the plan safe and executable?trajectory, predictions, vehicle limits, monitorsapproved trajectory or fallbackevery plan cycle
ControlWhat actuator command follows it?trajectory and vehicle statesteering, throttle, brake20-100 Hz

The boundaries are not universal. Some systems combine behavior and motion in MPC. Learned planners may output trajectories directly. The interface contract still needs to make these responsibilities explicit.


3. Route and Map-Level Planning

Route planning operates over a graph:

nodes/edges = lanes, lanelets, areas, aisles, zones, docking approach segments
costs       = distance, time, lane change, closure, congestion, risk, energy
constraints = vehicle type, one-way rules, clearance, access, ODD state

For road and airside applications, Lanelet2-style maps split physical geometry from traffic/routing interpretation. For warehouses and yards, the graph may be built from aisles, staging zones, charging stations, trailer bays, and one-way operational rules.

Route output should not be just a polyline. It should preserve semantic context:

  • lane or area IDs
  • regulatory elements and stop/yield lines
  • speed zones
  • keepout zones
  • docking or handoff waypoints
  • closure and temporary restriction metadata

4. Behavior Planning

Behavior planning turns route intent and world state into maneuver intent.

Common behavior states:

  • lane follow / corridor follow
  • stop at line or goal
  • yield / creep / proceed
  • lane change or aisle change
  • overtake or pass obstacle
  • merge
  • reverse
  • dock / undock
  • pull over / minimal-risk condition
  • wait for clearance

Behavior planning must arbitrate between:

  • traffic or site rules
  • predicted agent motion
  • operator or ground-control instructions
  • fleet coordination reservations
  • ODD restrictions
  • fallback and safety policies

For airside autonomy, behavior planning should explicitly encode aircraft priority, stand-entry rules, runway/taxiway clearance boundaries, jet-blast keepouts, FOD zones, marshalling/ground-control instructions, and mixed GSE right-of-way policy.


5. Motion Planning Families

FamilyExamplesStrengthRisk
Graph searchA*, Dijkstra, hybrid A*Reliable over maps and grids; easy to debug.Resolution and heuristic quality affect smoothness and compute.
State latticemotion primitives, lattice plannersEnforces vehicle kinematics by construction.Primitive library must cover required maneuvers.
Sampling-basedRRT, RRT*, PRMHandles high-dimensional spaces and complex constraints.Output may need smoothing; narrow passages can be hard.
Optimization-basedquadratic programming, nonlinear programming, MPCDirectly handles smoothness and constraints.Needs good initialization and solver timing bounds.
Frenet polynomialquintic/quartic lateral-longitudinal candidatesEfficient for lane/corridor following.Depends on a valid reference path and projection.
Kinodynamic planninghybrid A*, trajectory rollout, reachabilityAccounts for dynamics and controls.More compute and modeling effort.
Learned/generativediffusion, imitation, RL, VLA trajectory headsCan capture human-like interactions and multimodality.Requires safety envelope, validation, and OOD monitoring.

Most production-style systems are hybrids: route graph plus behavior state machine plus motion optimization or lattice/Frenet candidates plus independent trajectory validation.


6. Trajectory Generation

6.1 Path, Speed, and Trajectory

Definitions:

path:       geometric curve, usually x(s), y(s), yaw(s), kappa(s)
speed plan: velocity and acceleration as a function of time or arc length
trajectory: time-parameterized state sequence x(t), y(t), yaw(t), v(t), a(t)

A controller needs a trajectory, not merely a path. A path without timing does not tell the vehicle when to brake, how to respect jerk limits, or how to handle moving obstacles.

6.2 Common Representations

RepresentationUse
piecewise linear pathsimple route and low-speed initial plans
clothoid / curvature-continuous pathroad-like steering comfort
cubic or quintic splinessmooth interpolation and path smoothing
quintic polynomialboundary-conditioned position/velocity/acceleration trajectory
quartic polynomialvelocity keeping with free terminal position
Bezier / B-splinecompact smooth curves with control points
discrete trajectory samplescontroller interface and validation

6.3 Frenet-Style Candidate Generation

For corridor-following, a common pattern is:

  1. Project current state onto reference path.
  2. Sample target lateral offsets, speeds, and time horizons.
  3. Generate lateral and longitudinal polynomial candidates.
  4. Convert candidates back to Cartesian coordinates.
  5. Reject candidates that violate collision, curvature, acceleration, jerk, or route boundaries.
  6. Choose the lowest-cost feasible trajectory.

This is fast and interpretable, but it needs a good reference path and can struggle with open areas, unstructured aprons, tight docking, and complex multi-agent negotiation unless augmented.

6.4 Speed Planning

Speed planning enforces:

  • speed limits and advisory speeds
  • stop lines and goal stops
  • predicted obstacle interactions
  • curvature-limited speed
  • comfort acceleration and jerk
  • degraded-mode speed caps
  • safe-stop distance

Curvature-limited speed:

v_max(s) = sqrt(a_lat_max / max(|kappa(s)|, epsilon))

For airside and industrial operations, speed should also depend on personnel proximity, aircraft proximity, load state, surface condition, visibility, and local site policy.


7. Trajectory Validation and Runtime Assurance

Every candidate trajectory should be checked against at least:

CheckWhat it protects
collision and swept volumestatic and predicted obstacles
drivable area / lane / zone boundariesmap and ODD compliance
curvature and curvature ratesteering feasibility
velocity, acceleration, jerkcomfort and actuator limits
stopping distancesafe fallback margin
time validitystale plans and delayed state
rule compliancestop/yield/clearance/one-way restrictions
localization and perception uncertaintyrisk margins under uncertainty
controller feasibilitytracking limits and command saturation

Validation should produce a reason when it rejects a plan. That reason is needed for debugging, operator trust, fleet analytics, and safety-case evidence.


8. Practical Deployment Notes

8.1 Interface Contract

A planner output should include:

  • frame and timestamp
  • trajectory points with pose, velocity, acceleration, curvature
  • route/lane/zone IDs used
  • maneuver state
  • obstacle assumptions and prediction horizon
  • validity duration
  • fallback trajectory or stop point
  • planner status and rejection reasons

8.2 Replanning Policy

Use separate horizons:

  • long route/task horizon: seconds to minutes
  • behavior horizon: several seconds
  • motion horizon: 3-10 seconds for many vehicles, shorter for tight indoor/docking
  • control horizon: immediate actuation window

Avoid mode oscillation by using hysteresis, commitment windows, and explicit abort conditions for maneuvers such as lane changes, merges, docking, and clearance-required crossings.

8.3 Domain-Specific Notes

DomainDeployment note
Road AVUse map semantics and prediction uncertainty to avoid behavior oscillation at merges/intersections.
Indoor AMRModel one-way aisles, human work zones, narrow-passage reservations, and docking handoff.
Outdoor industrialInclude terrain, grade, traction, dust/weather, trailer swing, and fleet traffic rules.
AirsideTreat aircraft and clearance boundaries as first-class constraints, not just obstacles. Keep stand docking separate from corridor following.

9. Failure Modes and Risks

Failure modeSymptomMitigation
Layer mismatchRoute says proceed, behavior says yield, motion planner oscillates.Define ownership and priority between route, behavior, motion, and safety layers.
Infeasible trajectoryController saturates or cannot track.Validate curvature, acceleration, jerk, delay, and actuator limits before control.
Cost-weight brittlenessPlanner chooses uncomfortable or risky paths after small tuning changes.Use hard constraints for safety and regression tests for cost changes.
Reference path dependencyFrenet planner fails in open aprons, yards, docks, or unstructured zones.Switch to area planners, lattice/hybrid A*, or docking-specific planners.
Stale map or closureRoute enters blocked or newly restricted area.Use map versioning, dynamic overlays, and route invalidation.
Prediction overconfidencePlanner cuts through an agent's plausible future path.Use calibrated prediction uncertainty and contingency/fallback planning.
Mode oscillationVehicle creeps, stops, creeps, or toggles lane-change decisions.Add hysteresis, commitment, and clear abort criteria.
Learned planner OODEnd-to-end model proposes plausible but invalid motion.Keep rule, feasibility, and safety validation outside the learned proposal.
Late planController tracks stale trajectory.Timestamp plans, enforce age limits, and provide fallback stop trajectories.


Sources

Public research notes collected from public sources.